Paul-Emmanuel sotir \paul-emmanuel@outlook.com
We ran 3 different hyperparameter searches.
The last one (./hp_search_logs/hp_detect3.log) trains the
latest version of the model which performs better:
best_valid_loss=0.0313197 (didn't performed well due to early, buggy model
implementation) best_valid_loss=0.0039857
best_valid_loss=0.0031991See following visualizations for better understanding of hyperparameter search results on ball detection model:
summarize_hp_search(r'../hp_search_logs/hp_detect3.log', 'Ball detection 3')
summarize_hp_search(r'../hp_search_logs/hp_detect2.log', 'Ball detection 2')
summarize_hp_search(r'../hp_search_logs/hp_detect1.log', 'Ball detection 1')
We ran 2 different hyperparameter searches.
Even thought the first hyperparameter search did found a better model
(best_valid_mse=0.0005018) than the lastest hp search
(best_valid_mse=0.0003088), the latest model version seems to be more
promizing for further improvements during a full/regular training with more that 90 epochs. Indeed,
first/older hp search trials globaly didn't performed better and we can suspect that its best trial is
overfitting on validset much more than the best trial of the newer/latest hp search:
summarize_hp_search(r'../hp_search_logs/hp_forecast2.log', 'Ball position forecasting 2')
summarize_hp_search(r'../hp_search_logs/hp_forecast1.log', 'Ball position forecasting 1')